CoreML model is crashing on iPhone GPU, but not on iPhone CPU or macOS GPU

🐛 Describe the bug

A customer has reported a speech-to-text recognition runs correctly on CoreML on the macbook via pybindings, but crashes when run on the iPhone using the GPU compute unit:

/Library/Caches/com.apple.xbs/Sources/MetalPerformanceShadersGraph/mpsgraph/MetalPerformanceShadersGraph/Core/Files/MPSGraphTensorData.mm:203: failed assertion `shape.count = 0 != strides.count = 2'

The model does run on the iPhone when the compute unit is set to CPU_ONLY. A self-contained repro is provided here: https://github.com/chmjkb/canary-et-repro/blob/main/coreml_canary_repro.ipynb

I have confirmed the model runs on the macbook via pybindings, and by extracting the mlpackage from the *.pte file I have confirmed that it runs on the GPU on the macbook without crashing with the profile tool. I have also confirmed the mlpackage runs on the iPhone using CPU_ONLY with the profile tool, but crashes on the iPhone using CPU_AND_GPU.

Versions

NA

cc @kimishpatel @YifanShenSZ @cymbalrush

Start with the self-contained coreml_canary_repro.ipynb and reproduce the model on iPhone with CPU_ONLY versus CPU_AND_GPU. Compare the extracted mlpackage runs using the profile tool and the reported MPSGraphTensorData assertion; done means the iPhone GPU path no longer crashes while CPU and macOS behavior remain working.